If attention plays such a crucial role in the valuation process, could we utilize this information to improve choices? Nudging can be defined as an aspect of the choice architecture that alters people’s choices and subsequent wellbeing without compulsion, to improve “decisions about health, wealth and happiness” (Thaler and Sunstein, 2008).
As mentioned above, corroborating evidence suggests that stimulus values are computed by determining values and weights for an item’s attributes. All attributes are then integrated into an overall subjective value (Rangel, 2013). The attribute inte-gration process is thought to depend on the attention that is deployed among the at-tribute characteristics (Hare et al., 2011a). The attentional DDM suggests that visual saliency can interact with the comparator process and influence choices (Krajbich et al., 2010). Thus, attention can be shifted towards certain cues (e.g., nutrition labels), which should bias choices towards the attended option or attribute if the value is posi-tive (with opposite effects if its value is negaposi-tive, Fehr and Rangel, 2011). In case of dietary choice, the weighting of health (long-term) and taste (short-term) aspects is of importance. For example, if individuals put a large weight on short-term attributes, this may come at the expense of possible negative long-term health consequences.
35
A seminal study suggested that self-control in dietary choice depends on an individual’s ability to incorporate both health and taste information into the vmPFC value signal (Hare et al., 2009). In a follow-up study, Hare and colleagues (2011) ma-nipulated the attentional focus in dietary choices by instructing participants to focus on health, taste, or no particular attribute while choosing whether they wanted to eat a food item at the end of the experiment. On a behavioral level, they found that cues directing attention towards health features increased the weight on health attributes in their choices. In addition, healthy choices were correlated with the degree that health ratings reflected the vmPFC value signal. Also, health attention cues increased activi-ty in regions of the dlPFC, which in turn indirectly modulated the vmPFC (Hare et al., 2011a). Thus, attention manipulation towards health cues led to behavioral and neural changes resembling those of endogenous self-control (Hare et al., 2009, 2011a). Naturally, asking participants to direct their attention on choice attributes is not feasible in every-day life. Henceforth, it is of high interest to unravel potential means to improve goal-directed decision making by subtly influencing attribute atten-tion deployment, see studies 5.1 and 5.2. For instance, salient cigarette warnings, compared to text-based information, were shown to increase smoking cessation, pos-sibly due to increased attention on long-term health attributes (Borland et al., 2009;
White et al., 2008). The computational model described above would explain this outcome with increased weight on long-term attributes, and possibly decreased weight on short-term rewarding properties of the cigarettes, however, this has not been explicitly tested. As mentioned previously, so-called “nudges” could be used to improve consumer welfare (Thaler and Sunstein, 2008). Various factors, such as framing, attention, and saliency, have been shown to affect choics, but having said that, a unifying model accounting for these effects has not been employed. Most probably, many of the results could be explained by changes of subjective value com-putations via attention (Fehr and Rangel, 2011).
36
4 Computational modeling in value-based decision making
The previous chapters already introduced - in passing - various computational models of decision making. Computational models have been used in various do-mains of psychology, such as perceptual decision making (Ratcliff, 1978), classical conditioning (Rescorla and Wagner, 1972), and other forms of learning (Gluck and Granger, 1993). There is currently also a strong trend in the field of value-based deci-sion making to model the computations required for goal-directed decideci-sions (Pezzulo et al., 2014). Two important models in the field are therefore discussed in more detail in the following subchapters: Drift Diffusion Modeling as a model of stimulus value computation and comparison, as well as Dynamic Causal Modeling as a general com-putational model for inferring effective connectivity between brain regions. After an introduction of the methods, applications in the field of decision making research are provided.
4.1 Drift Diffusion Modeling
Recall that the standard economic literature has traditionally only used out-comes of decision making, that is, choice data, for theory development and hypothe-sis testing (“revealed preference approach” (Samuelson, 1938)). However, psycholo-gists and neuroeconomists have introduced formal models of the decision making process, which has proven to add important insights into the underlying mechanism.
For instance, RTs are closely related to preferences (Krajbich et al., 2014).
DDMs are models of decision making that “provide a mathematical frame-work to understand decisional processes” (Voss et al., 2015) by decomposing choice and reaction time data into distinct parameters that can be used to infer internal psy-chological processes (Voss et al., 2013a, 2015). Originally, the DDM has been em-ployed in perceptual tasks in which participants decide, for instance, which of two stimuli is brighter, or which of two numbers is larger. The DDM assumes continuous information sampling until sufficient information is gathered for one of the possible
37
outcomes (Ratcliff, 1978; Ratcliff and Smith, 2004). Response time distributions from the two decision thresholds are used to estimate a predefined parameter set, which may vary in complexity (Voss et al., 2015). The DDM offers an alternative to classical RT and choice analysis as parameter estimates can be employed to infer un-derlying cognitive processes (Metin et al., 2013; Philiastides and Ratcliff, 2013;
Ratcliff and Smith, 2004; Voss et al., 2013a).
In detail, the process of information sampling can be described by “a Wiener diffusion process” with a slope (i.e., drift (v), Voss et al., 2015) towards one of the boundaries and Gaussian noise (see Figure 7). The drift rate parameter is frequently
“interpreted as the speed of information uptake” (Voss et al., 2015). “The better the quality of evidence, the larger the drift rate toward the appropriate decision boundary, and the faster and more accurate the response” (Philiastides and Ratcliff, 2013). The two thresholds (“barriers“) at 0 and a represent the two alternative outcomes of the decision process; |a| denotes “the amount of information that separates both possible decisional outcomes” (Voss et al., 2015). Larger threshold separation leads to longer decision times but fewer errors (Metin et al., 2013; Voss et al., 2015). If one decision-al outcome is preferred, the starting point (z) is positioned further, i.e., biased, to one of the boundaries. Thus, if the starting point z is closer to one of the two thresholds, less information is needed to reach a decision (Voss et al., 2015). Extra-decisional processes, such as task preparation and stimulus encoding that take place before the comparison and decision phase, as well as motor execution after the decision process are mapped onto a single parameter, the non-decision component t0 (note the differ-ence to the parameter v, which relates to the processes during the comparison and decision phase and depends on the quality of evidence).
38
Figure 7: Simplified version of the Drift Diffusion Model: At the beginning of a trial, the information accumulation process starts at a starting point (z) and runs over time with a mean slope (drift, v), until a barrier (a, or 0) is hit. Reaction time distributions are presented above and below the decision thresholds. For simplicity, the non-decision time (t0) before and after the diffusion process is not shown. From Voss et al.
(2013).
While most DDM research has been done in the domain of perceptual deci-sion making (Heekeren et al., 2008; Ratcliff and Smith, 2004; Voss et al., 2004), re-cent studies have used the information accumulation process models to analyze eco-nomic decisions (Busemeyer and Townsend, 1993; Hare et al., 2011b; Krajbich et al., 2010, 2015; Milosavljevic et al., 2010; Towal et al., 2013). Various lines of research propose that using RTs along with choice data can improve preference predictions (Krajbich et al., 2014, 2015).
Perceptual DDM tasks usually rely on stimuli that are stochastic in nature (such as different brightness levels). In contrast, in value-based decision-making, the noisy brain representation of the choice options is thought to lead to stochastic choic-es (Glimcher, 2014; Krajbich et al., 2014). While perceptual decision making requirchoic-es very similar stimuli in order to compare different experimental tasks, economic deci-sions are based on value comparisons as a “common currency” (Brosch & Sander, 2013). For example, the DDM fit to a food choice experiment could quite accurately predict choices and RTs in social-preference experiments (Krajbich et al., 2015).
39
In a refinement of the original DDM, the economic decision DDM assumes that every decision involves a dynamic computation of a relative decision value (RDV) variable (Fehr & Rangel, 2011).
In detail, when a decision maker has to decide between two options x and y,
“the decision maker observes value signals xt and yt, randomly drawn from two dis-tributions with means ux and uy” (Krajbich et al., 2014). At time t, the individual ob-serves value signals and updates her RDV. The RDV evolves over time (see Figure 9).
Note that this model assumes that the decision maker cannot instantly access his/her preferences, but repeatedly samples from normal distributions and applies a sequen-tial likelihood ratio test (Krajbich et al., 2014). Because of the stochasticity of choice, there is the probability that individuals choose an option of lower subjective value (Fehr and Rangel, 2011). Choice “mistakes” result from too little samples, distribu-tions lying very close together (choice difficulty, see Figure 8), or lower barriers (for example due to a speed instruction).
Figure 8: The DDM in economic decision making assumes that decision makers can-not immediately access his/her preferences, but randomly sample from normal distri-butions around the true value signal. Here, it is assumed that the dotted lines are the true value signals, and the decision maker samples three times from each distribution (dots). Note that if the decision maker only samples once from each distribution, for example the right-most dot of the blue distribution, and the left-most dot of the red distribution, the outcome will be a “mistake” – that is, the left item will be chosen over the right item, although the right item is preferred.
40
In detail, the RDV, assuming that signals are drawn from normal distributions, evolves over time according to
Equation 9
where d is a parameter governing the speed of accumulation of the RDV. Triv-ially, the average rate at which evidence is accumulated relies upon on the difference in subjective value between the two options under consideration (Krajbich et al., 2014, 2015). Thus, when the absolute value difference is large, the RDV evolves with a steeper slope, compared to small value differences; see Figure 9. “Difficult” deci-sions, that is, decisions between roughly equally liked items (or decisions between equal amounts of moving dots in both directions in case of a perceptual Random Dot Motion task, for example), result in longer RTs.
Figure 9: The relative decision value (RDV) evolves over time until one of two deci-sion thresholds is reached. The slope is dependent on the value differences of item x and y (ux and uy). From Krajbich et al. (2014).
Note that from a classical economic perspective, decisions between items of equal value should be very fast, as time is a valuable resource and both options are equally utility-maximizing. Empirical evidence and the model demonstrates, and
pre-41
dicts, respectively, the opposite (Krajbich et al., 2014). The DDM could be fit to a vast-amount of perceptual and value-based decision-tasks (Fehr and Rangel, 2011;
Krajbich et al., 2015; Metin et al., 2013; Milosavljevic et al., 2010; Philiastides et al., 2010; Philiastides and Ratcliff, 2013; Rangel and Hare, 2010; Ratcliff and Smith, 2004; Schmitz and Voss, 2012; Voss et al., 2004, 2013a, 2013b).
Functional MRI and non-human electrophysiology data suggest that the putations described by the DDM are similar to the way the brain computes and com-pares value signals (Basten et al., 2010; Hare et al., 2011b; Heekeren et al., 2008;
Philiastides et al., 2006). Further refinements of the DDM include the multi-alternative DDM (Krajbich and Rangel, 2011), the attentional DDM (Krajbich et al., 2010, 2012), see chapter 3.2, and the neural DDM (Hare et al., 2011b).
Applications of DDM in value-based decision making 4.1.1
As outlined above, the DDM implies a suboptimal use of time, as decisions between items of equal value take the longest although both items are equally utility-maximizing. In one study, researchers analyzed whether it is possible to improve this suboptimal use of time by imposing a time limit on individuals’ choices. In a block without time limit, participants could take as much time as they wanted for each choice, but had to complete 100 choices in 150 seconds. Unreached choices led to random draws of choices. In a time-limit block, participants were asked to choose as quickly as possible whenever a choice took too long. If participants did not reach a decision within half a second after the onset of the prompt, the choice was drawn ran-domly. The time limit improved participants’ final choice surplus compared to no time limit, that is, the mean difference in value between the chosen and worst option on the screen (this difference is naturally higher when random, instead of preference-based, choices are made). The authors concluded that “DDMs yield new empirically validated insights into the potential sub-optimalities” of decision makers, which “can be mitigated with novel policy interventions” (Krajbich et al., 2014). In a similar line of research, participants were asked to make value-based decisions under low and high time pressure. The researchers found that the distance between the two decision thresholds (a), that is, the amount of information a participant requires before initiat-ing a response, significantly decreased under time pressure (Milosavljevic et al., 2010), which is in line with results in perceptual decision making tasks (Ratcliff and
42
McKoon, 2008; Voss et al., 2004). Philiastides & Ratcliff (2013) used brand labels of modulators of value in binary decisions between products. In two blocks, products were presented either with a brand label, or without one. In the block without labels, participants chose based on their subjective values. However, branding biased sub-jects’ decisions towards the more-preferred brand, which could be explained by changes in drift rate only. The study suggests that information on brands and subjec-tive preferences are integrated into an value signal in the decision making process (Philiastides and Ratcliff, 2013).
4.2 Dynamic Causal Modeling
Dynamic causal modeling (DCM) is a biologically plausible Bayesian frame-work “for inferring hidden”, that is, unobserved, “neuronal states from measurements of brain activity” (Stephan et al., 2009). In a standard mass univariate functional neu-roimaging analysis, one can use statistical parametric maps (SPMs) to localize differ-ences in brain activity between, e.g., different tasks, conditions or populations (Fris-ton et al., 1991). More specifically, one can use experimental manipulations (for ex-ample different tasks or a pharmacological intervention) to attribute significant re-gional or focal activations to the processes (sensorimotor or cognitive) manipulated in task A compared to task B, or under pharmaceutical A versus B (Friston et al., 1994).
This method only allows the analysis of direct experimental effects on each voxel, without permitting connections between nodes and their modulation (Friston et al., 2003). Psycho-Physiological Interaction (PPI) analysis, a type of functional connec-tivity analysis, is used to investigate “task-specific changes in the relationship be-tween activity in different brain areas” using regression analysis (O’Reilly et al., 2012). For example, one could analyze the statistical dependency between a region of interest and the rest of the brain during task A compared to task B. Importantly, PPI analysis does not make inferences on the direction of information flow (if region A influences region B, or region B influences region A, or both, or whether the influ-ence is mediated via another region; O’Reilly et al., 2012).
In contrast to functional connectivity analyses, DCM allows inferring effec-tive connectivity, that is, a causal relationship between brain areas, for example, whether activity in region A causally influences activity in region B, or whether
ex-43
perimental task A modulates the connectivity between brain region A and brain region B. In DCM, the brain is regarded “as a deterministic nonlinear dynamic system”, which is subject to experimental “inputs and produces outputs” (Penny et al., 2011).
More specifically, the input, that is, the experimental manipulation, is regarded as
“perturbation of neuronal dynamics”, which is distributed through coupled nodes (Penny et al., 2011). Thus, exogenous experimental stimuli evoke brain responses, which can influence network nodes: Inputs can a) influence state variables (neuronal activity) directly, or b) influence the coupling between nodes (Marreiros et al., 2010, Stephan et al., 2007). An example for a direct influence would be visual stimulation, while an example for an indirect influence would be attention (Friston et al., 2003).
To this end, a dynamic input-state-output model is employed. The inputs are usually explanatory variables (from a conventional design matrix as used in classical univariate fMRI analyses). The outputs are hemodynamic responses in the considered brain regions. DCM uses a forward model of how brain regions respond to experi-mental inputs (as neuronal activity cannot be directly measured using fMRI and is therefore inferred, Friston et al., 2003), see Figure 10.
Neuronal dynamics in nodes (z) are transformed into BOLD signals (y) via a hemodynamic response function (λ); DCM uses this forward model to estimate pa-rameters at a neuronal level, with the aim to maximize the similarity between predict-ed and estimatpredict-ed BOLD signal (Stephan et al., 2009). Given the neural state equation (see Figure 10), that is, the change in neural systems, the neural parameters can be expressed as partial derivatives of the endogenous connectivity (A-Matrix), modula-tory input (B-Matrix) and direct input (C-Matrix, Friston et al., 2003; Marreiros et al., 2008). It is important to note that “DCM does not assume temporal precedence” as a necessity for causality, as “the lag between neuronal activity and BOLD activation can theoretically vary across brain regions” (Ballard et al., 2011).
44
Figure 10. Summary of the state equation used in DCM: The dynamics of the brain system consisting of various nodes cannot be directly observed (orange boxes) and are “determined by experimental manipulations”. Driving inputs, such as visual stimulation, elicit local responses, which are propagated thorough the system accord-ing to intrinsic connections. These in turn can be altered by modulatory inputs, such as changes in task or attention. The integration of the state equation (green box) pro-duces predicted neural dynamics (z), “which enter a model of the hemodynamic re-sponse (λ) to give predicted BOLD rere-sponses” (y, “hemodynamic forward model”).
From SPM 8 Manual, Ashburner et al. (2010).
A ubiquitous question in any modeling approach is which model to select - in most modeling approaches, the decision is not solely made by comparing the relative fit of the alternative models, but also by accounting for the relative complexity, that is, the number of free parameters of the models (Ashburner et al., 2010). While more parameters improve model fit, the generalizability of a model often quite drastically decreases. Bayesian model selection (BMS) is employed for determining the most likely model among competing hypotheses about the mechanisms generating the ob-served data. “BMS is based on model evidence, which is the probability of obtaining a particular model, given the data” (Rosa et al., 2012). A model space with n nodes has 2n×n permutations of connections that can be turned on or off, which can be
addi-45
tionally modulated by experimental inputs (Friston and Penny, 2011; Rosa et al., 2012). DCM analyses with several nodes and only few prior hypotheses about the nature of the network hence lead to a combinatorial mass of all plausible models;
estimating all possible models in model space is for this reason computationally ex-pensive (Hillebrandt et al., 2013). In a relatively novel, more explorative model selec-tion procedure, that is, “post-hoc” Dynamic Causal Model selecselec-tion, it is possible to determine the best model out of all possible connection structures. Because only a single model is estimated, this procedure drastically reduces computation time, allow-ing to search huge model spaces (Rosa et al., 2012).
Post-hoc DCM can be performed in three steps (based on the appendix in Enax et al., 2015b, see original appendix for additional details):
a. Eigenvariate extraction
In a first step, one needs to extract each participant’s principal eigenvariate (pref-erably around an individual’s local maximum activation closest to the peak voxel identified in a second level group analyses) of at least two regions of interest (nodes) within a specific radius at a specific, relatively liberal, predefined threshold. Whenev-er no supra-threshold voxels in one or more of the nodes can be extracted, the partici-pant has to be excluded from further analyses
b. Specification of model space and estimation
Driving input and, if applicable, modulatory input, is specified by using regres-sors from a conventional GLM. A model is a “full model” in a sense that it incorpo-rates all plausible reciprocal fixed connections between and within the nodes of inter-est (A-Matrix), and, if applicable, their modulation by modulatory input (B-Matrix).
The driving input can enter one or n nodes (C-Matrix). The specified model can then be inverted, that is, estimated.
46 c. Post-hoc model selection
To explore all possible DCMs, a post-hoc model selection routine can be ap-plied (Friston and Penny, 2011; Rosa et al., 2012). The post-hoc search takes a subset of parameters with the least evidence, and searches over all reduced models within that subset (by turning connections “off”). With more than 8 parameters, the post-hoc routine implements a “greedy search” over all models formed by removing all permu-tations of eight parameters whose individual removal produces the smallest reduction in model evidence, resulting in 28 reduced models. All possible combinations of disa-bling these parameters are evaluated, the model with the greatest evidence is selected, and the steps are repeated until no more connections can be pruned (Crone et al., 2015; Friston and Penny, 2011; Rosa et al., 2012). Post-hoc routines were shown to yield results comparable to conventional DCM model selection procedures (Rosa et al., 2012).
Applications of DCM in value-based decision making 4.2.2
Making causal inferences about the structure of a neuronal network is highly attractive for decision neuroscientists.
For instance, Hare and colleagues (2011a) used DCM to examine the mecha-nism through which the activity in regions related to self-control (that is, three sub-regions of the dlPFC) modulated value signals in the vmPFC. Next to classical, bilin-ear DCMs, the authors allowed brain regions to influence the coupling strengths be-tween other brain regions (using so-called “non-linear” DCM). “These modulation parameters capture the degree to which changes in the activity of one region modulate the coupling between two other regions” (Hare, 2011a), independent of experimental stimuli. They found that two regions that were more active during blocks in which participants had to consider health attributes in their decision modulated the coupling between another region of the dlPFC (where activity was correlated with obtained health ratings) and the vmPFC (which correlated with the overall subjective value, Hare et al., 2011a). Hence, the authors provided a computational model of how an individual employs self-control in value-based decisions.
47
As alluded to in previous chapters, it is conjectured that whenever an individ-ual faces a choice between stimuli, the brain assigns values to each stimuli, compares then, and then activates a motor response to implement the decision. Thus, stimulus values need to be transformed into motor commands. In a study specifically address-ing the brain’s underlyaddress-ing computational model, Hare and colleagues found that the vmPFC, as expected, encoded stimulus values in a binary choice task. The authors found evidence that the vmPFC value signals are passed to regions in the intraparietal sulcus and dorsomedial prefrontal cortex, presumably stimulus values are compared.
The output of these “comparator regions” then modulated activity in the motor cortex at the time of decision. This modulation was choice-dependent, in that these regions increased connectivity with the left motor cortex, whenever the right option was cho-sen, and to the right motor cortex, whenever the left option was chosen (Hare et al., 2011b). Therefore, DCM provides a model of how the brain computes values, com-pares them, and executes the decision.
Another important issue in value-based decision making is of course how mo-tivation translates goals into actions. Ballard and colleagues (2011) analyzed where reward information enters the brain (“entering” as implied by the DCM framework), and how reward information modulates the mesolimbic reward system. Using DCM, they identified that goal-directed information enters the network in the dlPFC. Re-ward information predicting high (but not low) reRe-wards then increased the directed connection strength from the dlPFC to regions of the mesolimbic reward system (ven-tral tegmental area and nucleus accumbens), structures important for motivated be-havior. Thus, this study elegantly suggests a model how the dlPFC integrates reward information in a context-dependent manner and then implements goal-directed behav-ior by influencing the mesolimbic dopamine system (Ballard et al., 2011).
48
5 Published studies during the qualification phase: con-text and summary
As discussed in the previous chapters, various contextual factors, such as sali-ency, affect the choice process and choice outcomes. A unifying model accounting for such effects is missing. Most likely, observed effects of contextual cues on preference formation and choice behavior could be explained by changes in the computation of a subjective value via attention on the attribute or item (Fehr and Rangel, 2011). In four studies, I analyzed how contextual product cues change the valuation process and preferences on a behavioral and neural level. The studies are briefly summarized in the sections below and can be found using the citations below the header.
5.1 Nutrition labels influence value computations in the ventromedial prefrontal cortex
Published in: Enax, L., Hu, Y., Trautner, P., & Weber, B. (2015). Nutrition labels in-fluence value computation of food products in the ventromedial prefrontal cor-tex. Obesity, 23(4), 786–792. https://doi.org/10.1002/oby.21027
As introduced in chapter 2.3 and 3.4, the basic model of subjective value computations assumes that the value computed at each point of time is the weighted sum of stimulus attributes, and the weight of attributes depends on attention alloca-tion (Fehr and Rangel, 2011; Hare et al., 2011b; Rangel and Clithero, 2014), see Equation 1, Equation 7, and Equation 8. “Errors” in the computation process are thought to arise due to the inability of decision makers to take into account certain attributes, such as long-term health consequences (Fehr and Rangel, 2011, Hare et al., 2011). Policy-makers often want to positively influence choices and bias them to-wards choosing options with higher long-term benefits (Kable, 2014). These interven-tions may change the degree that certain attributes are employed in the computation of a decision value, thereby promoting healthier eating habits (Hawkes et al., 2015).
Nutrition labels have an important role in informing individuals about the health attributes food products (Sonnenberg et al., 2013). In this study, I used
nutri-49
tion labels as value modulators. In detail, I compared how a more salient, color-coded label, compared to an information-based numeric label influences product valuation on a behavioral and neural level. The study was based on two studies by Hare and colleagues (2009, 2011). As stated above, in the first study (2009), hungry partici-pants were asked to make dietary choices between food items, which varied in taste and health attributes. They found that activity in the vmPFC correlated with stimulus value signals. In self-controlling individuals, the value signal incorporated both health and taste attributes, while it only reflected taste attributes in non-self-controlling indi-viduals. They found that exercising self-control increased activity in the dlPFC, which in turn modulated the value signal in the vmPFC (Hare et al., 2009). In their follow-up study (2011), participants were asked to evaluate food items but the re-searchers exogenously manipulated the attention on certain attributes. Participants were asked to focus on health, taste, or no particular attribute. They found that the exogenous manipulation of attention towards health attributes increased activity in the dlPFC and changed the value signal in the vmPFC accordingly. As instructing subjects to focus on health attributes is not feasible in daily life, a more subtle way of increasing the integration of health attributes in the valuation process may be achieved with salient nutrition labels.
In this experiment, 35 healthy participants were instructed to valuate different food items, which were presented in combination with a salient, color-coded, or a numeric, information-based label. They engaged in a BDM auction (see chapter 2.4, incentivized WTP). I conjectured that red labeling on unhealthy items, compared to numeric labeling on unhealthy items, should decrease WTP and increase activation in regions implicated in self-control and response inhibition, that is, the dlPFC (Hare et al., 2009, 2011a; Horn et al., 2003; Simmonds et al., 2008). On the other hand, green labeling on healthy items, compared to numeric labeling on health items, should in-crease WTP and activation in regions implicated in delayed reward anticipation, such as the posterior cingulate cortex (Kable and Glimcher, 2007; McCoy et al., 2003).
The change in valuation should be reflected in the vmPFC. It is of course of interest to analyze the network that changes the valuation. For this reason, I investigated whether regions implicated in self-control and delayed reward expectation showed increased connectivity to the vmPFC at the time of valuation.
50
I found that salient labels indeed influenced the valuation process in the ex-pected direction, in that the subjective value for green-labeled products increased, while it decreased for red-labeled products, compared to numeric labels. The vmPFC correlated with the subjective value of the items across label types. The functional MRI data suggested similarities between endogenous self-control in dietary choice and red labeling, as it activated a region of the dlPFC and exhibited increased func-tional coupling (PPI) with the vmPFC. The posterior cingulate cortex, implicated in top-down attention and internal goal representation (Gusnard and Raichle, 2001;
Hopfinger et al., 2000) was more activated in response to green labeling and showed increased connectivity to the vmPFC valuation system, suggesting that green labeling may increase long-term reward expectations.
The attentional DDM introduced in chapter 3.2 predicts that exogenous changes of attention, for example via nutrition labels, changes valuation processes. In particular, the model suggests that it biases valuations in favor of the more salient option when its value is positive, and it should have opposite effects when its value is negative (Fehr and Rangel, 2011). Indeed, this is what I found: the green label, com-pared to the numeric counterpart, more saliently highlights a product’s (in this case positive) health attributes and increased the subjective value of the products. In con-trast, red labels, compared to the numeric label, highlights a product’s negative health attributes and decreased the subjective value. The neuroimaging data provide infor-mation on the mechanistic details. Due to the rather poor time resolution of functional MRI, what remains unclear is how exactly health and taste information are dynami-cally integrated in the decision process between items, and how the value comparison process is influenced by the more salient label, which can be investigated using dy-namic models of the choice process, such as the DDM, see study 5.2.